Research On Semi-Automatic Small Target Annotation Technology In Infrared Sequence Images

Infrared detection of small targets has been widely used in military application field,for example,defense,precision strike and battlefield intelligence and reconnaissance.With the development of infrared small target intelligent detection research,the demand for measured infrared small target data in real world is expanding.However,the efficient annotation methods for infrared small target datasets are limited,and the existing efficient labeling methods are difficult to apply to infrared data.There are some problems in the actual annotation operation,manual annotation is inefficient,which is a kind of limit in the development of infrared small target detection technique.Therefore,aiming at two common forms of infrared small targets(point target and extended target),this paper will carry out the study on the efficient semi-automatic annotation method of infrared small target datasets.The method proposed in this paper can improve the annotation speed with less manual intervention while ensuring the annotation quality of datasets.The main work and research achievements are as follows:In order to improve the efficiency of infrared point target labeling,a semi-automatic infrared point target annotation process is proposed in this paper based on humanmachine collaboration.In this process,the difference between the annotation results of adjacent frames can be used to select key frames,which can enable the annotators to locate the image frames that may have errors and manually annotate them quickly,thus realizing efficient annotation with human-machine collaboration.In view of low signal-to-noise and strong maneuverability of infrared point targets,the existing labeling methods cannot be used.This paper proposes an automatic annotation process of infrared moving point targets based on target enhancement and visual tracking.Firstly,the infrared sequence images in a continuous time are registered and background eliminated to enhance the target features;Then,visual tracking algorithm is used to locate the enhanced features efficiently and automatically;Finally,the target saliency map of a single frame image is obtained by the phase spectrum reconstruction,and then the target is located accurately.The experimental results show that the automatic labeling method for point objects in complex backgrounds proposed in this paper can improve the labeling efficiency and shorten the labeling cycle.Aiming at the problems that the infrared extended target is difficult to extract effective features and the application scene is complicated,resulting in frequent failure of automatic annotation.this paper studies the semi-automatic annotation method of infrared extended targets based on dynamic programming tracking.Firstly,this method uses the prediction results of support vector machine(SVM)to perform weighted adaptive fusion of multiple features,in order to provide conditions for the stable tracking of infrared small targets.Then,when using dynamic programming to find the least cost path,the constraint weight of the energy function is adjusted according to the jitter degree of the imaging platform to enhance the anti-jitter of the method.The experimental results show that the improved dynamic programming tracker can be applied to a variety of infrared small target detection scenes,and can achieve efficient and high-quality annotation with as little manual participation as possible.In this paper,a semi-automatic infrared small target annotation system is designed and implemented according to the specific needs of the “Aerospace Cup” data annotation practice.Firstly,this paper clarifies the functions of the annotation system based on user needs.Then,the functional modules are designed in detail to determine the operation process of each module.Finally,build an infrared small target semi-automatic annotation system,and the system functions are displayed.In practical application tests,the system can speed up the labeling speed,improve the annotation quality of datasets and reduce the burden of labeling personnel to a certain extent.